QUICK SEARCH:   [advanced] Author: Keyword(s):
Home     Help     Feedback     Subscriptions     Archive     Search     Table of Contents

This Article Summary Full Text Full Text (PDF) Alert me when this article is cited Alert me if a correction is posted Services Email this article to a friend Related articles in JEB Similar articles in this journal Similar articles in PubMed Alert me to new issues of the journal Download to citation manager Citing Articles Citing Articles via HighWire Citing Articles via Google Scholar Google Scholar Articles by Fauria, K. Articles by Collett, T. S. Search for Related Content PubMed PubMed Citation Articles by Fauria, K. Articles by Collett, T. S.

Learning speed and contextual isolation in bumblebees

Karine Fauria, Kyran Dale, Matthew Colborn and Thomas S. Collett^*

School of Biological Sciences, University of Sussex, Brighton BN1 9QG, UK

View larger version (13K): [in a new window]   Fig. 1. (A) Plan view of the arena. Areas marked grey or surrounded by dashed lines are close to patterns at the feeder and nest ends of the arena and are those within which hovering times were measured. Holes at the feeder and nest ends of the box were 3 cm in diameter. They were placed at a height of 22 cm from the floor and were separated horizontally by 25 cm. (B) The different patterns to which the bees were trained. B, blue; Y, yellow.

View larger version (28K): [in a new window]   Fig. 2. Performance of bees with 15 or 17 trials of training in stage 2. Top: training patterns in stages 1-3. B, blue; Y, yellow. Bottom: percentage correct choices of groups of eight bees plotted against trial number for the different stages. (A) Performance with 45° versus 135° gratings at the feeder and yellow versus blue at the nest in stage 2, and with 45° versus 135° at the feeder and 135° versus 45° at the nest in stage 3. In stage 2, bees were assisted in trials 1 and 2, so these trials are not plotted. (B) Performance with 45° versus 135° at both the feeder and the nest in stage 2, and with 45° versus 135° at the feeder and 135° versus 45° at the nest in stage 3. Bees were assisted for the first two trials of stage 1 and these are not plotted.

View larger version (25K): [in a new window]   Fig. 3. Hovering times of bees in the experiment illustrated in Fig. 2A. (A,B) Total hovering time at the feeder and nest ends. Total hovering time averaged across bees is plotted against trial number. (C,D) Relative hovering times at the feeder and nest ends spent in front of positive (+ve) or negative (-ve) patterns. Values are means ± S.E.M. (N=8).

View larger version (27K): [in a new window]   Fig. 4. Performance of bees with five or seven trials of training in stage 2. Conventions and arrangement are as in Fig. 2. (A) No stage 1 training. Bees in stage 2 were assisted in trials 1 and 2, so these trials are not plotted. (B) Twelve trials in stage 1, with bees assisted on the first two trials, which are not plotted. B, blue; Y, yellow.

View larger version (25K): [in a new window]   Fig. 5. Performance of bees without stage 2 training. Conventions and arrangement as in Fig. 2. (A) Four trials in stage 1 of training with bees assisted in trials 1 and 2, for which the data are not plotted. (B) Seventeen trials in stage 1 of training. The first two trials were assisted and the data are not plotted. B, blue; Y, yellow.

View larger version (33K): [in a new window]   Fig. 6. Hebbian learning network. (A) Architecture of network. Each configural unit receives weighted input from all contextual (C1, C2...Cn) and local (L1, L2...Ln) units through plastic links that are subject to Hebbian and anti-Hebbian reinforcement during training. Inhibitory connections between the configural units produce a `winner-takes-all' output. The most active unit inhibits the rest and excites the output node (O). (B) Training cycle in pseudo code. Learning rules in the body of the code are applied until the network responds correctly or the permissible number of training cycles (MAX_TRAIN_CYCLES) is exceeded. (C) Flow chart of comparison of sequential and simultaneous training, as outlined in the text. av., average; std., standard deviation.

View larger version (32K): [in a new window]   Fig. 7. Difference in number of learning trials for correct performance in context 2 (C2) between sequential (B) and simultaneous (A) training. Many different initial weight conditions are used. Histogram showing the difference in the number of trials required to reach errorless performance in C2. Sequential training was faster to the left of zero (B-A<0) and simultaneous training was faster to the right of zero (B-A>0). Each histogram shows data with a different learning rate and number (4, 10 or 20) of configural units (given at the top of each plot).

View larger version (21K): [in a new window]   Fig. 8. Trajectories of weight changes during sequential and simultaneous training. (A) The network of two configural units (U1 and U2) and the inputs to them (L1, C1) (L2, C2). The weights (w) of the bracketed inputs to both U1 and U2 have been summed to give the axes of weight space in panels B and C. (B) The initial values of wC2 +wL2 (indicated by S) for both U1 and U2 are close. In consequence, U1 and U2 compete for the control of C2 at the start of simultaneous training. This interference keeps the input weights to U1 within the dashed circle. Eventually, the conflict is resolved and the input weights to U1 and U2 diverge. Interference is avoided with sequential training because the weights of U1 and U2 have separated to the positions S1 by the end of the first stage of training. (C) When the initial weights of U1 and U2 are well separated, there is no interference with simultaneous training, and thus no benefit from sequential training.